Xenos as Physics Processor?

[expletive]

http://www.xbitlabs.com/news/video/display/20051005202331.html

Interesting. Has this been discussed before?

How might somehting like this be implemented in practice and what would be the performance implications?

J

 

[Acert93]

Yeah, a few times off and on. Not much detail is known on how practical it is now, but down the road maybe?

Interestingly Xenos and R520 seem to be slowly opening this door. We had a thread/poll on this a week or two ago. It will be an interesting development.

It is my opinion that this type of behavior is planned for DX10 and that eventually GPUs will be taking up these parallel processing tasks, like physics. They are well suited for it, and the overlap between the markets (gamers who want pretty graphics, gamers who want nice physics, gamers who buy new GPUs) is pretty good I see the death of the PPU already, and its name is the GPU. Only so much room in a case/buyers budget, the GPU consolidating just one more task (like it did vertex processing from the CPU) makes it more vital and important. GPUs are a big part of Vista, I expect in the long run they will rival (gasp!) the CPU on many levels, primarily because it is much more scalable and benefits from parallelization.

 

[expletive]

Yeah, a few times off and on. Not much detail is known on how practical it is now, but down the road maybe?

Interestingly Xenos and R520 seem to be slowly opening this door. We had a thread/poll on this a week or two ago. It will be an interesting development.

It is my opinion that this type of behavior is planned for DX10 and that eventually GPUs will be taking up these parallel processing tasks, like physics. They are well suited for it, and the overlap between the markets (gamers who want pretty graphics, gamers who want nice physics, gamers who buy new GPUs) is pretty good I see the death of the PPU already, and its name is the GPU. Only so much room in a case/buyers budget, the GPU consolidating just one more task (like it did vertex processing from the CPU) makes it more vital and important. GPUs are a big part of Vista, I expect in the long run they will rival (gasp!) the CPU on many levels, primarily because it is much more scalable and benefits from parallelization.

Really great summary, thanks.

J

 

[Acert93]

TechPowerUp.com R520 Slides

"New Performance Architecture... Ideal for Physics and Data Parallel Processing"

Tomshardware on R520

The real innovation however is just an idea at this time, but promises to have a major impact on the industry, if brought to reality. Heye mentioned that ATI plans to open the hardware architecture of the X1000 to allow third party developers to write non-graphics-related applications to run on the graphics processor. The company calls this feature "dynamic load balancing."

Compared to a Pentium 4 CPU, which delivers a floating point performance of 12 GFLOPs and a bandwidth of just under 6 GByte per second, a graphics processor is a calculation monster: According to ATI, an X1800 XT chip reaches 83 GFlops and 42 GByte per second. The full performance of a graphics may not always be needed - especially in dual-graphics environments - and users will be able to relocate processing power to other applications. According to ATI, these applications could include scientific applications such as fluid dynamics, but also entertainment-related functions such as physics or 3D audio processing. Similar features have been demonstrated by academic projects in the past on ATI and Nvidia platforms, but dynamic load balancing as described by ATI officials promises a whole new use of graphics processors.

The company expects GPU specific third-party API's to become common within a few years - with one of the most promising being physics processing: ATI believes that graphics chips provide enough power to cover the features that are currently promoted by Ageia. If ATI's vision comes true, Ageia's idea for physics board for every gaming PC may

Dave's Xenos article also covers MEMEXPORT and some other features that could be very useful in "GPGPU" type processing. A small snip:

For instance, this is probably the first time that general purpose physics calculation would be achievable, with a reasonable degree of success, on a graphics processor and is a big step towards the graphics processor becoming much more like a vector co-processor to the CPU.

We have discussed this a little, off and on. As for real world implications I think some time is needed. But the recent ATI Toy Movie maybe (?) a small implication of where some of this is going... exciting times.

 

[expletive]

We have discussed this a little, off and on. As for real world implications I think some time is needed. But the recent ATI Toy Movie maybe (?) a small implication of where some of this is going... exciting times.

This Toy Movie, can I assume the X360 could render this in realtime with the same quality as well? I hadnt seen this before, wow.

J

 

[Acert93]

This Toy Movie, can I assume the X360 could render this in realtime with the same quality as well? I hadnt seen this before, wow.

J

It did Ruby at 30fps (someone mentioned in another thread in the GPU section that the 360 even got out of sync on bland scenes because it was processing quickly... probably the biproduct of a 2 week application port). My guess is if the R520 is doing it then yes. Although not a PC GPU, Xenos seems quite capable of handling the types of tasks R520/G70 are doing in the demo movies. From a "specs on paper" standpoint Xenos seems quite capable.

Of course this is all a guess based on paper specs and Ruby. Don't it don't mean much.

 

[Nemo80]

every DX9 card could do that in theory, that's no news. Guess why it's never been done up to now and why we now got PPU cards instead...

 

[Titanio]

If you have an interest in some of the technical/implementation detail of current work in this area, these links may be of interest:

http://download.nvidia.com/developer/presentations/2005/SIGGRAPH/ClothSimulationOnTheGPU.pdf

http://download.nvidia.com/developer/presentations/2005/I3D/I3D_05_GPGPU.pdf

http://www.gpgpu.org/s2004/slides/harris.SimulationGPU.ppt

www.gpgpu.org is generally a very good resource too, of course. GPU Gems 2 has dedicated sections on Simulation and GPGPU too, if you wanted to look it up.

 

[blakjedi]

Acert is funny. Slick way of referencing your own poll/thread. Well done! LOL


I was REALLY excited by the possibilities explored in that thread but still havent gotten around to watching the demos.. tonight then....

 

[Jawed]

Xenos should be quite a bit faster than X1800XT, so yeah it would run this demo pretty easily.

There should be a PDF appearing soon with an overview of how the demo was made.

http://www.beyond3d.com/forum/showthread.php?t=24254

Or you can rummage through the slides here:

http://www.noticias3d.com/articulo.asp?idarticulo=527&pag=10

There's some physics running on the GPU in this demo, e.g. the water droplets running down the shop window.

Jawed

 

[Nemo80]

Xenos should be quite a bit faster than X1800XT, so yeah it would run this demo pretty easily.

My bet, the opposite is true

Lower clockspeed, less efficient "pipes" and much lower transistor count.

 

[expletive]

every DX9 card could do that in theory, that's no news. Guess why it's never been done up to now and why we now got PPU cards instead...

Yes but wouldnt GPUs utilizing USAs be much more suited for these types of multipurpose calculations?

J

 

[3roxor]

Xenos should be quite a bit faster than X1800XT..

Based on what..? Looking at the 360 games it draws a different picture..

 

[expletive]

Based on what..? Looking at the 360 games it draws a different picture..

I dunno, seeing what epic has done with Gears of War in such a short time with final hardware I'm not sure i agree. Aside from other very impressive titles, the new Bioware and Silicon KNights stuff shown at X05 was pretty impressive as well.

J

 

[Master-Mold]

Based on what..? Looking at the 360 games it draws a different picture..

That comment makes no sense.

You can give a world class artist the best supplies in the world and tell him to paint a picture.

However if you only give him half his normal time and make him use unfamiliar tools the final product may be less than desired, but that doesnt mean his paint and canvas were bad.

 

[Acert93]

Acert is funny. Slick way of referencing your own poll/thread. Well done!

There's some physics running on the GPU in this demo, e.g. the water droplets running down the shop window.

Pretty amazing slides Technically this is VERY impressive.

Lower clockspeed, less efficient "pipes" and much lower transistor count.

7800GTX has a lower clock than the X1800XL, ditto the AMD 2.0GHz (3200) compared to a P4 3.2GHz. Clock speed is irrelevant when you are talking about a different architecture.

You are wrong on the efficent "pipes". Xenos pipes may not be as robust as say, G70, but they are definately more effecient due to utilization.

Per transistor count, Xenos has 257M transistors for logic. When you consider that R520/G70 have 1.) video processing engines (which are well over 20M transistors the last time I saw stats on them from a year ago) and 2.) have hardwired fixed functionality to support OLDER APIs the gap in transistors closes QUICKLY. You are looking at G70 being ~275M without PureVideo compared to Xenos' 257M. And that is not counting

Put in the context of past GPUs, this is not much. R420 had 160M transistors compared to NV40s 222M. I believe R300 had ~110M compared to 135M for NV30.

So when you consider the difference between a dedicated Console part with no legacy hardwired fixed function support or video processing Xenos is QUITE favorable in the transistor count.

And looking at Architecture and how things are implimented it is a very intelligent and effecient design.

And to harken back to a real work example, Xenos ran Ruby *fine* with a mere 2 week port. Ruby was designed for R520 on an x86 platform, so the fact it ran well on Xenos is a good indication Xenos is not some lame duck.

 

[Titanio]

Yes but wouldnt GPUs utilizing USAs be much more suited for these types of multipurpose calculations?

In terms of the computational model, they're SM3.0+. Really SM3.0 upwards + better precision everywhere is the significant breakthrough in terms of GPGPU imo - though yeah, not necessarily just the min DX9 spec (SM2.0 etc.).

 

[Acert93]

In terms of the computational model, they're SM3.0+. Really SM3.0 upwards + better precision everywhere is the significant breakthrough in terms of GPGPU imo - though yeah, not necessarily just the min DX9 spec (SM2.0 etc.).

Yes, and as we advance on the GPU will take on more typical CPU like features like branching and flow control. These will benefit GPGPU tasks as well as shading.

Exciting stuff. SLI/Crossfire that can be either used for physics or 2x the graphics performance? Now THAT would lure me in. Not a big SLI fan overall. Had a Voodoo2, ended up just getting a better GPU when my VooDoo2 could not keep up. Better features + Performance were too much to overcome. But the idea of upgrading to a new GPU and using the old one as a physics card =

 

[Titanio]

Yes, and as we advance on the GPU will take on more typical CPU like features like branching and flow control. These will benefit GPGPU tasks as well as shading.

Exciting stuff. SLI/Crossfire that can be either used for physics or 2x the graphics performance? Now THAT would lure me in. Not a big SLI fan overall. Had a Voodoo2, ended up just getting a better GPU when my VooDoo2 could not keep up. Better features + Performance were too much to overcome. But the idea of upgrading to a new GPU and using the old one as a physics card =

Yeah, I've wondered the same. The bus needs to get faster and faster though, if the CPU and GPU are to start working together on stuff I think. As is, the GPU could work independently on isolated stuff, I guess. There's also the numerical stability issue between the GPU/CPU to overcome from that perspective too (thinking about physics/simulation here).

 

[Nemo80]

And to harken back to a real work example, Xenos ran Ruby *fine* with a mere 2 week port. Ruby was designed for R520 on an x86 platform, so the fact it ran well on Xenos is a good indication Xenos is not some lame duck.

I don't know which ruby you are talkng about, but the one i know ran on my 9800 at about 15 fps, so i wouldn't expect anything less.

Anyways, it's DX9 all over, with likely zero CPU utilization, so porting a simple demo using the same API is not a big deal.